Seedance 2.5: Reference-to-Video vs. Video Editing

How Seedance 2.5 decides between reference-to-video, video editing, and video extension from your assets and prompt, why a task gets misread as editing, and how to write requests that behave predictably.

Short Answer

Seedance 2.5 has no dedicated “reference-to-video” switch. As soon as content contains a reference video, the model reads the intent of your prompt to decide whether the task is reference-to-video, video editing, or video extension:

• The prompt changes the source video (add, remove, modify, replace, keep something unchanged) → video editing
• The prompt continues the source video forwards or backwards (extend, continue) → video extension
• The prompt only borrows a character, motion, or style from the assets to shoot a new clip → reference-to-video

Once a task is classified as editing, ratio must be adaptive and duration must be -1. Passing a specific aspect ratio or duration returns a 400, usually mentioning TaskTypeConstraint.

This page applies to Seedance 2.5 (doubao-seedance-2-5-260628) only. The Seedance 2.0 family has no video editing or video extension tasks, so the issue does not arise there.

The Core Difference: Does the Asset End Up in the Output?

	Reference-to-video	Video editing
Role of the asset	Semantic reference only: appearance, motion, camera work, style, voice	The source video is the base of the output; the model adds, removes, or changes things on top of it
Output aspect ratio	Your choice (any of the seven ratio values)	Locked to the source video; ratio must be adaptive
Output duration	Your choice (duration from 4 to 30)	Locked to the source video; duration must be -1
Source video requirements	None	Must be 4–30 seconds long; under 20 seconds works best
Typical prompt	”A dancer on the beach, character from image 1, choreography referencing video 1"	"Replace the person in video 1 with image 1”, “Remove the background music from video 1”

A quick test: can you still see the source video itself in the output? If yes, it is editing (or extension). If only its “feel” carries over, it is reference-to-video.

Video extension is similar to editing: the aspect ratio is locked to the source video, but you can still set the duration.

How the Model Decides

The decision takes two steps: first the asset role, then the prompt.

Task type	Asset condition	Prompt trigger words (provider docs)	ratio	duration
Reference-to-video	At least one reference_image / reference_video / reference_audio	No editing or extension intent	Any	Any
Video editing	At least one reference_video	edit video, add, remove / delete, modify / replace / change to	Must be adaptive	Must be -1
Video extension	At least one reference_video	extend forwards / backwards, continue	Must be adaptive	Any
First frame / first & last frame	role is first_frame / last_frame	Independent of the prompt	Must be adaptive	Any

The trigger-word list is not exhaustive. The model judges meaning, not exact words. Phrases such as “keep everything in the video unchanged”, “upscale video 1 to HD”, or “keep the outfits the same” are not on the list, but they all describe processing the original footage, so they can be classified as video editing too.

A request with reference images but no reference video is never classified as editing or extension. You only need to watch for this when a reference_video is present.

A Real Case

This request wanted to “upscale a video” while also setting 4:3 and 15 seconds:

{
  "model": "doubao-seedance-2-5-260628",
  "ratio": "4:3",
  "duration": 15,
  "resolution": "1080p",
  "content": [
    { "type": "text", "text": "Upscale reference video 1 to HD, keep all elements in the video unchanged, keep the outfits unchanged" },
    { "type": "video_url", "role": "reference_video", "video_url": { "url": "asset://asset-xxxx" } }
  ]
}

It was rejected immediately with a 400:

The parameters `ratio` and `duration` specified in the request are not valid.
Seedance identified your task as video editing based on your prompt. ...
Issues: [0] `ratio` must be `adaptive`. [1] `duration` must be -1.

“Upscale” plus “keep unchanged” means working on the original footage, so the model classified it as editing, and editing tasks do not accept a custom aspect ratio or duration. Since this request really is an edit, the correct fix is:

"ratio": "adaptive",
"duration": -1

After the change, the output follows the source video’s aspect ratio and duration.

Can a Parameter Force Reference-to-Video?

No. In 2.5, omni_reference_task_type accepts only three values:

Value	Effect
auto (default)	The model decides from the assets and the prompt
edit	Declares video editing; editing constraints are validated at submission
extend	Declares video extension; extension constraints are validated at submission

There is no “reference-to-video” value. And edit / extend only validate earlier; they do not force the task type. If the declared type differs from what the model infers from the prompt, the task still fails with InvalidParameter.TaskTypeMismatch.

In short, the prompt decides the task type. Parameters can only go along with it.

Three Reliable Approaches

Output size doesn't matter

The general-purpose setup recommended in the provider docs. Whenever reference assets are present, send the following. No matter which subtask the model picks, the request will not fail on parameter constraints.

"omni_reference_task_type": "auto",
"ratio": "adaptive",
"duration": -1

The trade-off: when the task is classified as reference-to-video, the model chooses the duration (10 seconds or more in our tests), and the cost varies with it. Not recommended for cost-sensitive workloads.

Fixed aspect ratio and duration

To set ratio and duration yourself, the model must classify the task as reference-to-video. That comes down to the prompt:

• Describe the assets as something to reference, borrow from, or imitate, and say what is referenced (motion, camera work, style, character look)
• Focus on the new scene you want, not on what to do to the source video
• Avoid wording like add / remove / modify / replace / change to, “keep … unchanged”, “upscale / restore”, or “extend / continue”

Likely read as editing	Rewritten as reference-to-video
Replace the person in video 1 with the girl in image 1	The girl from image 1 runs along the beach, motion and camera work referencing video 1
Keep the scene in video 1 and add a cat	A cat walks past a street corner, in the street style of video 1
Upscale video 1, keep the outfits unchanged	A new runway shot featuring the character look and outfits from video 1

Retry in code

When prompts come from end users and you cannot control them, catch this 400 and resubmit once. The error is returned at submission, so no task is created, and 400 parameter errors are not billed.

import os
import requests

URL = "https://api.apiyi.com/seedance/api/v3/contents/generations/tasks"
HEADERS = {
    "Authorization": f"Bearer {os.environ['APIYI_API_KEY']}",
    "Content-Type": "application/json",
    "Accept-Encoding": "identity",
}

def submit(payload: dict) -> str:
    resp = requests.post(URL, headers=HEADERS, json=payload, timeout=300)
    if resp.status_code == 400 and "TaskTypeConstraint" in resp.text:
        # Classified as edit / extend / first-frame: release ratio and duration, then resubmit
        payload = {**payload, "ratio": "adaptive", "duration": -1}
        resp = requests.post(URL, headers=HEADERS, json=payload, timeout=300)
    resp.raise_for_status()
    return resp.json()["id"]

After the retry you no longer control the aspect ratio or duration. If your product must guarantee output specs, return the error to the user and ask them to rephrase instead of retrying silently.

When the Error Appears

Scenario	When it fails	Error code
edit / extend declared explicitly, parameters violate the constraints	400 at submission	InvalidParameter.TaskTypeConstraint
Omitted or auto, model classifies as editing, parameters violate the constraints	Provider docs say the task fails asynchronously, but it can also return a 400 at submission (as in the case above)	InvalidParameter.TaskTypeConstraint
Declared type differs from the model’s classification	Task fails after it starts running	InvalidParameter.TaskTypeMismatch

The two errors need different fixes: for TaskTypeConstraint, change the parameters; for TaskTypeMismatch, change the prompt (or set omni_reference_task_type back to auto).

Cost Notes

• A video-editing output is as long as the source video, not as long as you wanted. A 25-second source is billed at roughly 25 seconds.
• For tasks with a reference video, the input video’s frames are also converted into billable tokens. Longer, higher-resolution sources cost more.
• With duration: -1 on a reference-to-video task, the model picks the length, which may be longer than expected.

Checking the source video’s length before submitting avoids most surprises. To verify the actual charge of a task, see How do I look up a Seedance video’s real cost by task_id?.

Related Docs

Video Generation API

Task types vs. parameter constraints, full request parameters

Seedance 2.0 / 2.5 Overview

Differences between 2.5 and 2.0, full editing and extension usage